Vehicle-road cooperation signal control method based on large model Agent

Abstract

The invention discloses a vehicle-road cooperation signal control method based on a large model Agent, relates to the technical field of intelligent traffic control, and aims to solve the problems of signal control response lag and insufficient V2X scheme prediction capability in the prior art. The method comprises the following steps: S1, constructing a V2X collaborative architecture driven by a large model Agent, wherein the V2X collaborative architecture comprises a sensing Agent, a decision Agent, an execution Agent and a communication Agent; s2, constructing a knowledge graph database containing various traffic event coping strategies to realize event-strategy-influence dynamic association; s3, deploying a lightweight model to a roadside edge calculation unit, and generating a signal phase scheme through the lightweight model; and S4, polling traffic flow sudden change indexes, and performing decision impact evaluation to formulate a dynamic response mechanism. By constructing a V2X collaborative architecture driven by a large model Agent, dynamic traffic flow millisecond-level decision, multi-target global optimization and emergency scene adaptive control are realized, and the traffic efficiency of an intersection and the system robustness are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent traffic control technology, specifically to a vehicle-road cooperative signal control method based on a large model agent. Background Technology

[0002] With the acceleration of urbanization, traffic congestion has become a key issue restricting urban development. Traditional traffic signal control systems mainly rely on fixed timing schemes or adaptive control based on simple induction coils (such as SCOOT and SCATS systems), lacking an agent system with environmental interaction, collaborative decision-making, and online evolution capabilities. This results in poor system flexibility and easily leads to problems such as empty lanes at intersections and excessively long queues.

[0003] In recent years, the development of vehicle-to-everything (V2X) technology has provided new ideas for traffic signal control, as illustrated in publication CN116156457A. Through vehicle-to-infrastructure (V2I) and vehicle-to-vehicle (V2V) communication, real-time vehicle location, speed, and steering intention data can be obtained. However, existing V2X signal control schemes are mostly based on preset rule engines (such as priority traffic rules and green wave coordination algorithms). Such rule systems have the following three limitations: lack of global coordination, single-intersection optimization cannot solve regional congestion, and multi-intersection coordination only supports simple linkage (such as fixed phase difference), making dynamic adjustment difficult; weak predictive ability, as rule engines cannot model complex spatiotemporal evolution patterns of traffic flow (such as platoon dispersion effect and intersection conflict chain propagation); and insufficient scenario generalization, requiring manual customization of rules for special scenarios such as ambulance priority and tidal flow lanes, resulting in poor scalability.

[0004] In summary, existing technologies have not yet solved the three core problems of real-time response to dynamic traffic flow, multi-objective global optimization, and adaptive response to sudden scenarios. Summary of the Invention

[0005] This invention addresses the problems of delayed signal control response and insufficient predictive capability of V2X schemes in existing technologies, and proposes a vehicle-road cooperative signal control method based on a large model agent, which significantly improves the traffic efficiency and system robustness of intersections.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a vehicle-road cooperative signal control method based on a large model agent, comprising the following steps: S1, Construct a large-scale model agent-driven V2X collaborative architecture, including perception agent, decision agent, execution agent and communication agent; S2, construct a knowledge graph library containing various traffic incident response strategies to achieve dynamic association between events, strategies, and impacts; S3, deploy a lightweight model to the roadside edge computing unit, and generate a signal phase scheme through the lightweight model; S4 polls traffic flow change indicators and performs decision reassessment to develop a dynamic response mechanism.

[0007] This technical solution deeply embeds a large model agent into the real-time control link, and achieves millisecond-level dynamic decision-making closed loop and autonomous response in complex scenarios with "dynamic knowledge base + millisecond-level inference". It breaks through the three major bottlenecks of traditional solutions: slow response, weak prediction ability and poor scenario adaptability. The results are in line with reality.

[0008] The present invention is further configured such that: after the perception agent acquires the roadside collection data, it performs data preprocessing on the collection data to form structured traffic data, and uses the upgraded spatiotemporal graph convolutional network to perform millisecond-level prediction of vehicle trajectories within a set range of the intersection to generate a traffic flow situation matrix.

[0009] In this technical solution, vehicle status information is mainly obtained by integrating millimeter-wave radar, video detectors, and on-board OBU in the roadside unit.

[0010] The present invention is further configured such that: the decision agent is equipped with a lightweight model, which receives traffic flow situation matrix, historical signal control scheme, historical queue length and sudden events, receives natural language instructions through structured prompt word templates, and outputs instructions containing phase number, green light duration, priority indicator and confidence score.

[0011] In this technical solution, the decision agent can also autonomously trigger confidence assessment. When the confidence level is less than 0.85, a three-stage fault tolerance mechanism is triggered: first, the historical similar scenario solution library is searched (cosine similarity ≥ 0.9), then the reinforcement learning online optimization module is started, and finally the traffic police platform is handed over to manual management.

[0012] The present invention is further configured such that: the execution agent sends instructions to the signal controller, and pre-transmits candidate phase schemes to the buffer of the signal controller before the green light phase ends, and monitors the execution status in real time with a hardware-level watchdog circuit. When the execution agent detects that the instruction is lost or abnormal, it will automatically switch to the previous valid scheme.

[0013] The present invention is further configured such that step S2 includes: It adopts multi-level ontology modeling to cover various scenarios in six major categories: traffic accidents, meteorological disasters, infrastructure failures, special events, traffic control, and abnormal behavior. Each event is associated with three-dimensional influencing factors. The Agent event response chain forms a closed-loop decision flow by autonomously perceiving environmental events, intelligently matching knowledge graph strategies, and triggering the execution of instruction sets.

[0014] In this technical solution, the manual review mechanism is also designed as a collaborative intervention interface for the Agent system.

[0015] The present invention is further configured such that the lightweight model construction process is as follows: through model pruning, quantization compression and knowledge distillation techniques, the pre-trained large language model is compressed into a lightweight model within a set maximum value.

[0016] The present invention is further configured such that: the upgraded spatiotemporal graph convolutional network includes a spatial graph convolutional layer, a temporal convolutional layer, and a spatiotemporal interaction module; the created graph structure is input into the upgraded spatiotemporal graph convolutional network, and the input of the spatiotemporal graph convolutional network is used to generate a traffic flow situation matrix, which includes eight core indicators: lane flow, speed, density, steering ratio, queue length, headway, number of conflict points, and frequency of rapid acceleration and deceleration.

[0017] The present invention is further configured such that the created graph structure includes treating each vehicle as a node and dynamically generating an adjacency matrix based on the relative positions of the vehicles, wherein the adjacency matrix includes the latitude and longitude coordinates of the same lane, adjacent lanes and other lanes.

[0018] The present invention is further configured such that step S4 includes: the dynamic response agent polls traffic flow mutation indicators in real time, and uses a refined reward function to perform millimeter-level decision reassessment, wherein the reward function is a weighted function of vehicle delay rate, priority vehicle passage ratio, low-speed fuel consumption index and number of conflict points.

[0019] The present invention is further configured such that step S4 also includes an adaptive optimization mechanism, wherein each intersection agent encrypts and uploads the decision gradient after differential privacy processing daily, and the cloud control center aggregates and generates global model incremental parameters to achieve a preset percentage of monthly model iteration optimization rate.

[0020] In this technical solution, a fault rollback logic is set up based on the above. When the decision confidence is less than 0.7 for multiple consecutive times, the system automatically switches to the historical best solution library to ensure the robustness of the system in extreme scenarios.

[0021] The present invention can bring the following beneficial effects: This application presents a vehicle-road cooperative signal control method based on a large model agent. By constructing a V2X cooperative architecture driven by a large model agent, it achieves millisecond-level decision-making for dynamic traffic flow, multi-objective global optimization, and adaptive control for sudden scenarios, thereby improving the traffic efficiency and system robustness of intersections. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating a vehicle-road cooperative signal control method based on a large model agent, as described in this application.

[0023] Figure 2 This is a diagram illustrating the collaborative implementation architecture of a vehicle-road cooperative signal control method based on a large model agent, as described in this application.

[0024] Figure 3 This is a schematic diagram illustrating the knowledge graph construction of a vehicle-road cooperative signal control method based on a large model agent, as proposed in this application. Detailed Implementation

[0025] Example 1 This embodiment proposes a vehicle-road cooperative signal control method based on a large model agent. By deploying a cluster of large model agents on the roadside, an intelligent agent cooperative network for autonomous perception, decision-making, and execution is constructed to achieve dynamic optimization control of traffic signals and improve traffic efficiency and safety.

[0026] refer to Figure 1 , Figure 2 as well as Figure 3 It includes the following steps.

[0027] Step S1: First, construct a large model agent-driven V2X collaborative architecture, which includes a perception agent, a decision agent, an execution agent, and a communication agent.

[0028] In this embodiment, each of the above-mentioned Agents achieves millisecond-level data interaction through a standardized interface.

[0029] The perception agent has the functions of data collection, preprocessing, and prediction.

[0030] The perception agent can acquire roadside data, preprocess the data to form structured traffic data, and finally use an upgraded spatiotemporal graph convolutional network to predict vehicle trajectories within a set range at intersections in milliseconds, generating a traffic flow situation matrix.

[0031] In this embodiment, the collected data mainly includes: trajectory data of the roadside unit (RSU) integrated millimeter-wave radar (positioning error ≤ 0.3m), target data of the video detector (target recognition rate ≥ 95%), and vehicle status information (including position, speed, and driving intention) of the on-board unit (OBU).

[0032] The data preprocessing includes data cleaning, format standardization, and spatiotemporal alignment, ultimately resulting in structured traffic data, which is then uploaded to edge computing nodes at a fixed frequency. In this embodiment, the fixed frequency is 20Hz.

[0033] In this embodiment, the upgraded Spatiotemporal Graph Convolutional Network (ST-GCN) is used to perform millisecond-level prediction of vehicle trajectories within a range of 50-200 meters from the intersection to obtain a dynamic traffic flow situation matrix.

[0034] The upgraded spatiotemporal graph convolutional network includes spatial graph convolutional layers, temporal convolutional layers, and a spatiotemporal interaction module. The created graph structure is input into the upgraded spatiotemporal graph convolutional network, and the input of the spatiotemporal graph convolutional network is used to generate a traffic flow situation matrix, which includes eight core indicators: lane flow, speed, density, steering ratio, queue length, headway, number of conflict points, and frequency of rapid acceleration and deceleration.

[0035] In more detail, the implementation process of the upgraded Spatiotemporal Graph Convolutional Network (ST-GCN) in traffic flow prediction includes the following sub-steps.

[0036] Step 101, dynamic trajectory extraction, includes trajectory data from millimeter-wave radar, target data from video detectors, and vehicle status information from the onboard OBU; in this embodiment, the frequency is 20Hz.

[0037] Step 102: Set a spatiotemporal window. The spatial range is a fan-shaped coverage area with a radius within a set range centered on the intersection. The time window is set to a sliding window length of T. A vehicle dynamic truncation strategy is adopted. A maximum vehicle processing threshold N is set. When there are more than N vehicles in the range, vehicles are filtered according to emergency vehicles > distance from the intersection.

[0038] In this embodiment, the spatial range is a fan-shaped coverage area with a radius of 50 to 200 meters centered on the intersection. The time window is set with a sliding window length of T=3s. A vehicle dynamic truncation strategy is adopted, and the maximum processing threshold N=50 vehicles is set. When there are more than N vehicles in the range, the vehicles are filtered according to the order of emergency vehicles > distance from the intersection.

[0039] Step S103, creating a graph structure, which includes: treating each vehicle as a node and dynamically generating an adjacency matrix A based on the relative positions of the vehicles. The adjacency matrix A includes the latitude and longitude coordinates of the same lane, adjacent lanes, and other lanes.

[0040] More specifically, when elements A(i,j) in the adjacency matrix A are in the same lane: A(i,j) is twice the value of e. -n , where n is the square of the distance between the latitude and longitude coordinates of vehicle i and vehicle j, divided by 2σ².

[0041] When an element A(i,j) in the adjacency matrix A is in an adjacent lane: A(i,j) is e -n , where n is the square of the distance between the latitude and longitude coordinates of vehicle i and vehicle j, divided by 2σ².

[0042] When the element A(i,j) in the adjacency matrix A is in another lane: A(i,j) is 0.5 times e. -n , where n is the square of the distance between the latitude and longitude coordinates of vehicle i and vehicle j, divided by 2σ².

[0043] Wherein, the latitude and longitude coordinates of vehicle i are represented as (xi, yi), the latitude and longitude coordinates of vehicle j are (xj, yj), and σ is the spatial attenuation coefficient; in this embodiment, the spatial attenuation coefficient is taken as 10 meters.

[0044] Step S104: Construct a spatiotemporal graph convolutional network, which mainly includes spatial convolutional layers, temporal convolutional layers, and a spatiotemporal interaction module.

[0045] Step S105: Predict the target based on the constructed spatiotemporal graph convolutional network, predict the trajectory coordinates of each vehicle at multiple time points in the future period; in this embodiment, predict the trajectory coordinates of each vehicle at 60 time points in the next 3 seconds.

[0046] Step S106 generates a traffic flow situation matrix, which includes eight core indicators: lane flow (number of vehicles passing through a certain section of a lane per unit time, unit: vehicles / hour), speed (average speed of vehicles in a lane, unit: kilometers / hour), density (number of vehicles per unit length of lane, unit: vehicles / km), steering ratio (percentage of vehicles turning left, going straight, and turning right at each approach lane, unit: %), queue length (queue distance of vehicles waiting to pass behind the stop line, unit: meters), headway (time difference between two consecutive vehicles passing through the same section in the same lane, unit: seconds), number of conflict points (number of points where trajectories may intersect), and frequency of sudden acceleration and deceleration (number of events of sudden deceleration (deceleration > 2.5 m / s²) or sudden acceleration (acceleration > 1.5 m / s²) per unit time).

[0047] In this technical solution, a corresponding lightweight model is set up inside the decision agent. The lightweight module can receive the traffic flow situation matrix, historical signal control scheme, historical queue length and sudden events output from step S1. It receives natural language instructions through structured prompt word templates and outputs instructions containing phase number, green light duration, priority indicator and confidence score.

[0048] The aforementioned lightweight model is specifically designed to compress a pre-trained large language model into a lightweight model within a set maximum value through model pruning, quantization compression, and knowledge distillation techniques.

[0049] In this embodiment, a lightweight model is deployed that has undergone structured pruning (removing 80% of redundant attention heads), INT8 quantization, and knowledge distillation (teacher model LLaMA-13B) compression. This model can receive traffic flow situation matrices, historical signal control schemes, historical queue lengths, and sudden events, such as emergency vehicle communication needs and temporary traffic control data. Subsequently, it receives natural language instructions based on structured prompt word templates, such as "East-West saturation 0.92, Northbound 120-meter ambulance requesting priority passage," and finally outputs JSON instructions containing phase number, green light duration, priority flag, and confidence score.

[0050] The decision agent can autonomously trigger confidence assessment. In this embodiment, when the confidence level is <0.85, a three-stage fault tolerance mechanism is triggered: first, the historical similar scenario solution library is searched (cosine similarity ≥0.9), then the reinforcement learning online optimization module is started, and finally the traffic police platform is handed over to manual management.

[0051] The execution agent can send instructions to the signal controller and pre-transmit candidate phase schemes to the buffer of the signal controller before the green light phase ends. Combined with the hardware-level watchdog circuit, the execution status is monitored in real time. When the execution agent detects that the instruction is lost or abnormal, it will automatically switch to the previous valid scheme.

[0052] More specifically, decision instructions are sent to the signal controller in real time via a low-latency bus (transmission delay ≤ 2ms). In this embodiment, candidate phase schemes are pre-transmitted to the buffer of the signal controller 3 seconds before the green light phase ends. Combined with a hardware-level watchdog circuit to monitor the execution status in real time, the execution agent will automatically switch to the previous valid scheme when it detects that the instruction is lost or abnormal, ensuring that the control interruption time is ≤ 100ms, thereby realizing the dynamic loading of phase schemes.

[0053] For the communication agent, in this embodiment, a redundant network is built based on the C-V2X PC5 interface, and multi-link aggregation technology (RSU-signal dedicated line + 5G-Uu public network backup) is adopted to achieve an end-to-end transmission latency of ≤20ms. A lightweight communication protocol is set up, and JSON commands are binary compressed (compression rate ≥80%). The communication agent uses a dynamic channel allocation algorithm to avoid co-channel interference, and the packet loss rate is controlled below 0.1% in densely populated road areas.

[0054] Step S2: Construct a knowledge graph library containing various traffic incident response strategies to achieve dynamic association between events, strategies, and impacts.

[0055] More specifically, multi-level ontology modeling is adopted to cover various scenarios in six categories: traffic accidents (e.g., rear-end collisions, rollovers), meteorological disasters (e.g., heavy fog FOG-04, heavy rain RAIN-12), infrastructure failures (e.g., traffic light power outages, road collapses), special events (e.g., large-scale events, emergency drills), traffic control (e.g., temporary road closures, major security operations), and abnormal behaviors (e.g., pedestrians running red lights, vehicles driving against traffic). Each event is associated with three-dimensional influencing factors (physical / spatial / temporal). The Agent event response chain forms a closed-loop decision flow by autonomously perceiving environmental events, intelligently matching knowledge graph strategies, and triggering the execution of instruction sets. At the same time, the manual review mechanism is designed as the collaborative intervention interface of the Agent system.

[0056] Step S3: Deploy the lightweight model to the roadside edge computing unit and generate a signal phase scheme through the lightweight model.

[0057] In this embodiment, a pre-trained large language model (such as LLaMA-7B) is compressed to a lightweight version of less than 50MB and deployed on the roadside edge computing unit through model pruning, quantization compression and knowledge distillation techniques.

[0058] The decision agent receives real-time data input from multiple sources, including traffic flow matrix, historical traffic data, emergency information, and priority passage requests. It proactively generates phase control instructions through semantic understanding, which include the current phase number, green light duration (accurate to the second), candidate next phase sequence, and special priority flags, while also adding a confidence score of 0-1.

[0059] The decision agent can autonomously trigger confidence assessment. When the confidence level is ≥0.85, it directly sends the decision to the signal controller for execution. Otherwise, it triggers the dynamic optimization module based on reinforcement learning or calls the historical solution library to ensure that the decision is completed within 10ms delay. At the same time, it uses a federated learning framework to achieve differential privacy encryption update of model parameters across intersections to continuously improve adaptability to complex scenarios.

[0060] Step S4: Poll traffic flow mutation indicators and conduct decision reassessment to develop a dynamic response mechanism.

[0061] More specifically, the dynamic response agent polls traffic flow change indicators in real time (e.g., lane occupancy rate change rate ≥15% / second, queue length exceeding the threshold of 150 meters, emergency vehicle approach distance <200 meters), and uses a refined reward function to perform millisecond-level decision reassessment. The reward function is a weighted function of vehicle delay rate, priority vehicle passage ratio, low-speed fuel consumption index, and number of conflict points.

[0062] The reward function R is the sum of several impact factors, including the first impact factor, the second impact factor, the third impact factor, and the fourth impact factor.

[0063] Among them, the first influencing factor is the product of the difference between 1 and the vehicle delay rate and the first empirical weighting coefficient; the second influencing factor is the product of the proportion of priority vehicles and the second empirical weighting coefficient; the third influencing factor is the negative of the product of the low-speed fuel consumption index and the third empirical weighting coefficient; and the fourth influencing factor is the negative of the product of the number of conflict points and the fourth empirical weighting coefficient.

[0064] Among them, the first, second, third, and fourth empirical weight coefficients are all dynamically optimized through machine learning. The vehicle delay rate is the ratio of the average vehicle delay time to the free-flow time within the current signal cycle, which reflects traffic efficiency. The priority vehicle passage ratio is the proportion of special vehicles with priority, such as ambulances, fire trucks, and buses, passing through within a unit cycle, which reflects emergency rescue efficiency. The low-speed fuel consumption index is the converted value of fuel consumption for vehicles with speeds <5km / h, which reflects the degree of environmental protection. The number of conflict points is the number of potential risk points of vehicle trajectory intersection conflicts within the prediction cycle, which reflects safety.

[0065] Step S4 also includes an adaptive optimization mechanism, in which each intersection agent encrypts and uploads the decision gradient after differential privacy processing daily, and the cloud control center aggregates and generates global model incremental parameters to achieve a preset percentage of monthly model iteration optimization rate.

[0066] In this embodiment, the adaptive optimization module relies on the federated learning framework. Each intersection agent encrypts and uploads the decision gradient after differential privacy processing every morning. The cloud control center aggregates and generates global model incremental parameters to achieve a monthly iteration optimization rate of ≥12.5%. Fault rollback logic is set up so that when the decision confidence is <0.7 for 5 consecutive times, it automatically switches to the historical best solution library to ensure the robustness of the system in extreme scenarios.

[0067] To address the shortcomings of existing technologies in real-time response to dynamic traffic flow, multi-objective global optimization, and adaptive control in unexpected scenarios, this embodiment constructs a V2X collaborative architecture driven by a large-model agent. This architecture enables millisecond-level decision-making for dynamic traffic flow, multi-objective global optimization, and adaptive control in unexpected scenarios, improving intersection traffic efficiency and system robustness. By deeply embedding the large-model agent into the real-time control link, a "dynamic knowledge base + millisecond-level inference" approach is used to achieve millisecond-level dynamic decision-making closed loops and autonomous responses in complex scenarios. This overcomes the three major bottlenecks of traditional solutions: slow response, weak predictive capabilities, and poor scenario adaptability, with practical results.

[0068] Example 2 A vehicle-road cooperative signal control method based on a large model agent includes the following steps: Step S1: Construct a large-scale model agent-driven V2X collaborative architecture, including a perception agent, a decision agent, an execution agent, and a communication agent; Step S2: Construct a knowledge graph library containing various traffic incident response strategies to achieve dynamic association between events, strategies, and impacts; Step S3: Deploy the lightweight model to the roadside edge computing unit and generate a signal phase scheme through the lightweight model; Step S4: Poll traffic flow mutation indicators and conduct decision reassessment to develop a dynamic response mechanism.

[0069] Based on the above embodiment 1, this embodiment will be further described in detail.

[0070] For details, please refer to Figure 1 and Figure 2 In a certain city, a three-layer collaborative architecture was constructed. The perception layer is responsible for collecting real-time location and speed data of vehicle-mounted OBUs and traffic flow and accident information data of roadside RSUs. The edge layer adopts triple optimization technology and deploys an agent decision-making module to achieve a 200ms response time. The cloud layer constructs a knowledge graph library of a four-dimensional relational network. Some case details of the knowledge graph library are shown in Table 1. Table 1. Some examples of knowledge graph databases .

[0071] As a preferred implementation, at a certain moment, when an ambulance enters the 200-meter range of an intersection about 300 meters from a hospital, the roadside sensing agent captures its location, speed (82 km / h, with warning lights and sirens activated) and driving intention (going straight through the intersection) in real time using millimeter-wave radar and video detectors. At this time, the real-time signal phase information shows that there are 10 seconds left on the green light for going straight, and the queue lengths at the east entrance are 120 meters for left turns, 95 meters for straight turns, and 60 meters for right turns. The spatiotemporal graph convolutional network detects that there are 3 vehicles that may obstruct passage for going straight east.

[0072] After receiving the real-time traffic flow matrix and the ambulance priority request, the decision agent generates instructions based on semantic templates using a lightweight large model: extend the east-west straight-ahead green light to 15 seconds (from 10 seconds), and display "Ambulance allowed, no rushing" on the north-south guidance screen, with a corresponding confidence level of 0.9. The instructions are directly issued, and the execution agent transmits them to the traffic signal controller via a low-latency bus, extending the east-west straight-ahead green light to 15 seconds. The 5G-V2X broadcast reaches 47 connected vehicles, of which 3 actively changed lanes to yield. The intersection camera provides real-time feedback on the ambulance's passage, confirming that it did not slow down throughout the journey (passing the intersection in 2.8 seconds).

[0073] After the event, the system automatically records the effect of the decision (ambulance delay 0 seconds, average delay of other vehicles increased by 2.1 seconds) and updates the model parameters through federated learning. The knowledge graph is simultaneously updated with new cases, linked to the "emergency vehicle priority" strategy, for use in similar scenarios later.

Claims

1. A vehicle-road cooperative signal control method based on a large-scale agent model, characterized in that, Includes the following steps: S1, Construct a large-scale model agent-driven V2X collaborative architecture, including perception agent, decision agent, execution agent and communication agent; S2, construct a knowledge graph library containing various traffic incident response strategies to achieve dynamic association between events, strategies, and impacts; S3, deploy a lightweight model to the roadside edge computing unit, and generate a signal phase scheme through the lightweight model; S4 polls traffic flow change indicators and performs decision reassessment to develop a dynamic response mechanism.

2. The vehicle-road cooperative signal control method based on a large model agent according to claim 1, characterized in that, After acquiring the collected data, the perception agent performs data preprocessing to form structured traffic data. It then uses an upgraded spatiotemporal graph convolutional network to perform millisecond-level prediction of vehicle trajectories within a set range at intersections, generating a traffic flow situation matrix.

3. A vehicle-road cooperative signal control method based on a large model agent according to claim 1 or 2, characterized in that, The decision agent is equipped with a lightweight model. The lightweight model receives traffic flow situation matrix, historical signal control schemes, historical queue lengths and emergencies, receives natural language instructions through structured prompt word templates, and outputs instructions containing phase number, green light duration, priority indicator and confidence score.

4. The vehicle-road cooperative signal control method based on a large model agent according to claim 3, characterized in that, The execution agent sends instructions to the signal controller. Before the green light phase ends, it pre-transmits candidate phase schemes to the buffer of the signal controller. Combined with a hardware-level watchdog circuit, it monitors the execution status in real time. When the execution agent detects that an instruction is lost or abnormal, it will automatically switch to the previous valid scheme.

5. A vehicle-road cooperative signal control method based on a large model agent according to claim 1, 2, or 4, characterized in that, Step S2 includes: It adopts multi-level ontology modeling to cover various scenarios in six major categories: traffic accidents, meteorological disasters, infrastructure failures, special events, traffic control, and abnormal behavior. Each event is associated with three-dimensional influencing factors. The Agent event response chain forms a closed-loop decision flow by autonomously perceiving environmental events, intelligently matching knowledge graph strategies, and triggering the execution of instruction sets.

6. The vehicle-road cooperative signal control method based on a large model agent according to claim 3, characterized in that, The lightweight model construction process is as follows: through model pruning, quantization compression, and knowledge distillation techniques, the pre-trained large language model is compressed into a lightweight model within a set maximum value.

7. The vehicle-road cooperative signal control method based on a large model agent according to claim 2, characterized in that, The upgraded spatiotemporal graph convolutional network includes a spatial graph convolutional layer, a temporal convolutional layer, and a spatiotemporal interaction module. The created graph structure is input into the upgraded spatiotemporal graph convolutional network, and the input of the spatiotemporal graph convolutional network generates a traffic flow situation matrix, which includes eight core indicators: lane flow, speed, density, steering ratio, queue length, headway, number of conflict points, and frequency of rapid acceleration and deceleration.

8. The vehicle-road cooperative signal control method based on a large model agent according to claim 7, characterized in that, The created graph structure includes treating each vehicle as a node and dynamically generating an adjacency matrix based on the relative positions of the vehicles. The adjacency matrix includes the latitude and longitude coordinates of the same lane, adjacent lanes, and other lanes.

9. A vehicle-road cooperative signal control method based on a large model agent according to claim 1 or 2, characterized in that, Step S4 includes: the dynamic response agent polls traffic flow change indicators in real time, and uses a refined reward function to perform millimeter-level decision reassessment. The reward function is a weighted function of vehicle delay rate, priority vehicle passage ratio, low-speed fuel consumption index and number of conflict points.

10. A vehicle-road cooperative signal control method based on a large model agent according to claim 1 or 2, characterized in that, Step S4 also includes an adaptive optimization mechanism, in which each intersection agent encrypts and uploads the decision gradient after differential privacy processing daily, and the cloud control center aggregates and generates global model incremental parameters to achieve a preset percentage of monthly model iteration optimization rate.

Images